From Claims to Compliance: Evaluating Verifiability in AI System Disclosures via Policy-Grounded Analysis

Model cards and evaluation reports inform AI governance, but the verifiability of their claims remains unclear. We analyze 412 safety, evaluation, and alignment claims from 68 Hugging Face model cards and 24 HELM reports using an OECD-aligned three-point scale. Only 38.1% of claims are well-supported and 20.1% unsubstantiated, with safety claims least verifiable (26.3% well-supported, 28.1% unsubstantiated). Bootstrapped intervals (33.5%–42.8%) reveal persistent gaps between disclosure and verifiability, motivating policy-aligned auditing standards.